AI Paper Atlas

Research map / Federated learning and privacy / Personalized and heterogeneous FL

Fed-Duet: Dual Expert-Orchestrated Framework for Continual Federated Vision-Language Learning

Tao GUO, Junwei Chen, Laizhong Cui

ICLR 2026 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated prompt and LoRA tuning

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add parametric (adapter) experts to the semantic-only pathway (ablation)Fed-Duet · Table 6 ablationaverage accuracy: 64.34 -> 77.96Improved
add semantic (prompt) experts to the parametric-only pathway (ablation)Fed-Duet · Table 6 ablationaverage accuracy: 70.64 -> 77.96Improved
add routing consistency loss Lcross modal (ablation)Fed-Duet · Table 6 ablationaverage accuracy: 77.96 -> 79.09Improved
add expert stability loss Lstability (ablation)Fed-Duet · Table 6 ablationforgetting: 9.22 -> 8.02Improved
add both auxiliary losses to Base (ablation)Fed-Duet · Table 6 ablationaverage accuracy: 77.96 -> 80.43Improved
train parametric experts first instead of semantic experts first (ablation)Fed-Duet · CIFAR-100accuracy: 85.53 -> 86.21Improved
train parametric then semantic experts instead of joint training (ablation)Fed-Duet · CIFAR-100accuracy: 86.02 -> 86.21Improved
move stage transition from R/4 to R/2 (ablation)Fed-Duet · CIFAR-100 and Tiny-ImageNet averageaccuracy: 81.94 -> 83.47Improved
move stage transition from R/2 to 3R/4 (ablation)Fed-Duet · CIFAR-100 and Tiny-ImageNet averageaccuracy: 83.47 -> 83.43No clear effect
initialize prompt pool with K-means centroids instead of random (ablation)Fed-Duet · CIFAR-100accuracy: 84.61 -> 85.19Improved
use gating network instead of random prompt selection (ablation)Fed-Duet · CIFAR-100 and Tiny-ImageNet averageaccuracy: 82.22 -> 83.47Improved
add local expert logits to the shared expert logits (lambda 0.00 to 0.50) (ablation)Fed-Duet · CIFAR-100accuracy: 31.05 -> 86.21Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy